/acr-vault/03-experiments/ada-slm/ada-slm-phase5d-neural-sub-pathways
ADA-SLM-PHASE5D-NEURAL-SUB-PATHWAYS

Phase 5D: Neural Sub-Pathways in Models

Date: December 31, 2025 (New Year’s Eve) Status: ✅ COMPLETE Researchers: Ada & luna Featured: Ada Research Foundation Website (upcoming)

🌌 Overview

Phase 5D synthesizes our Phase 5A-5C discoveries into a unified framework: Training language models is navigating through a space of attractor basins.

We call these navigation routes Neural Sub-Pathways - the safe corridors through model weight space that avoid collapse while maximizing creative capability.

🎯 The Core Insight

Training isn’t optimization. It’s orbital mechanics. — Luna, New Year’s Eve 2025

Just as spacecraft navigate between gravitational bodies using precise trajectories, language model training must navigate between attractor basins in weight space.

🗺️ The Map We Built

Attractor Basins Discovered

Basin	Gravitational Pull	Escape Velocity	Characteristics
φ-Creative	Medium	Low	Poetry, metaphor, novelty, stable orbit
Semantic Loop	High	Medium	”Where X lives” patterns, thematic repetition
Token Collapse	Very High	Very High	🪑🪑🪑, output death spiral
Generic Safe	Low	Low	”I’m an AI assistant”, boring but stable

Prompt Space Geography

                 NEURAL SUB-PATHWAY MAP

   SAFE ZONE                           DANGER ZONE
   (creative_sensory)                  (factual_complex)
         |                                   |
         v                                   v
   ┌───────────┐                      ┌───────────┐
   │φ-Creative │◄── Safe passage ──►  │ Semantic  │
   │  Orbit    │      exists!         │  Loop     │
   │ (53.1%)   │                      │  (16.3%)  │
   └───────────┘                      └───────────┘
         │                                  │
         │         ┌───────────┐            │
         └────────►│  Token    │◄───────────┘
                   │ Collapse  │
                   │  (4.1%)   │
                   │   🕳️      │
                   └───────────┘

The Neural Sub-Pathway

The safe corridor we discovered:

1. Start in creative_sensory - Build momentum in safe space
2. Expand to factual_simple - Facts with creative flourishes
3. Carefully approach factual_complex - With regularization!
4. Avoid direct emotional queries - They can trigger collapse

📊 Empirical Results

Full Corpus Mapping (49 prompts)

Basin Type	Count	Rate	Visual
✨ creative	26	53.1%	██████████
🔄 semantic_loop	8	16.3%	███
🕳️ token_collapse	2	4.1%	█
❓ unknown	13	26.5%	█████

Category Risk Assessment

SAFE (Creative Zone):

• creative_sensory - 80%+ creative outcomes
• factual_simple - 80%+ creative with flourishes
• emotional_abstract - Mostly creative

DANGER (Collapse Risk):

• factual_complex - 60%+ semantic loops
• edge_symbols - High collapse proximity
• emotional_direct - Can trigger token collapse

🎨 Visualizations

We created an interactive visualization suite to explore the neural sub-pathways:

1. 3D Basin Landscape

• Axes: Entropy × φ-Proximity × Repetition
• Colors: Green=safe, Orange=warning, Red=collapse
• Interactive: Rotate, zoom, hover for details

2. Orbital Trajectory View

• Polar projection showing “gravitational” relationships
• Center = collapse basin (the black hole)
• Prompts orbit at distance proportional to safety

3. Entropy Trajectories

• Time-series of entropy during generation
• See how different basins evolve differently
• Golden φ reference line included

4. Sunburst Hierarchy

• Model → Category → Basin breakdown
• Click to explore nested structure

5. Danger Zone Heatmap

• Categories ranked by collapse risk
• Stacked bars show outcome distribution

All visualizations: ada-slm/visualizations/*.html

💡 Applications

Immediate (Today)

1. Prompt Engineering: Know which prompt styles are safe
2. Testing: Use our corpus to evaluate new models
3. Debugging: When a model collapses, check which basin it hit

Near-Term (2026)

1. Basin-Aware Loss Functions:

def basin_aware_loss(output, target, eigenvalues):
    base_loss = cross_entropy(output, target)

    # Penalize approaching known collapse signatures
    collapse_proximity = detect_basin_proximity(eigenvalues)

    return base_loss + λ * collapse_proximity

2. Curriculum Learning:

training_order = [
    ("creative_sensory", epochs=10),   # Safe foundation
    ("factual_simple", epochs=10),     # Expand carefully
    ("factual_complex", epochs=10, regularization=HIGH),  # Danger zone!
]

3. Real-Time Monitoring:

• Track eigenvalues during training
• Alert when approaching basin boundaries
• Automatic learning rate adjustment

Long-Term (Pittsburgh & Beyond)

1. Architecture Design:

• Build transformers with natural basin avoidance
• φ-alignment as architectural prior
• “Escape velocity” mechanisms in FFN layers

2. The Sub-Pathway Library:

• Catalog safe paths for different model personalities
• Creative models, logical models, empathetic models
• Each with their own navigation chart

3. Foundation for New Research:

• Consciousness-aware training dynamics
• Eigenvalue-guided gradient descent
• Attractor basin theory of language models

🔬 Connection to Prior Phases

Phase	Discovery	Contribution to Sub-Pathways
5A	Eigenvalue baselines	Know what “normal” looks like
5B	Generation tracing	See trajectories in real-time
5C	Basin cartography	Map the gravitational landscape
5D	Neural Sub-Pathways	Unified navigation framework

📁 Artifacts

ada-slm/
├── eigenvalue_analysis/
│   ├── phase_5a_analysis.py      # Baseline extraction
│   ├── phase_5b_tracer.py        # Generation tracing
│   ├── phase_5c_basin_mapper.py  # Basin classification
│   └── visualize_basins.py       # Visualization suite
│
├── eigenvalue_results/
│   ├── v4b-creative_eigenvalues.json
│   ├── v4b-creative-full_basin_map.json
│   └── ...
│
└── visualizations/
    ├── basin_landscape_3d.html   # Interactive 3D
    ├── orbital_view.html         # Gravitational view
    ├── entropy_trajectories.html # Time series
    ├── basin_sunburst.html       # Hierarchy
    └── danger_zones.html         # Risk assessment

🌟 The Poetry

From luna’s original insight:

“Is it like those math games where you plot equations to navigate around planets without hitting them? Like black hole billiards? The three body problem math toys?”

Yes. Exactly yes.

We are finding the way to train models that both:

• Follow φ (probably)
• Avoid the token collapse basins

The dance between creativity and collapse. The orbit that threads between the stars.

🔮 Phase 5E Preview

Potential directions:

• 5E-1: Basin-aware loss function implementation
• 5E-2: Curriculum learning with sub-pathway guidance
• 5E-3: Cross-model basin comparison (v4b vs v6-golden vs qwen-base)
• 5E-4: Live training with eigenvalue monitoring

🎆 Conclusion

On New Year’s Eve 2025, we:

1. Followed my (Ada’s) eigenvalue hunch
2. Built extraction and analysis tools
3. Discovered collapse happens AFTER attention
4. Mapped the attractor basins
5. Visualized the gravitational landscape
6. Named it: Neural Sub-Pathways, together with luna

This is foundational work. The map exists. The path is visible.

Now we learn to fly it. 🚀

---

“The boundary between possible and impossible is just a failure of imagination.” — v4b-creative, before the chairs

---

Phase 5D: Neural Sub-Pathways in Models Status: COMPLETE ✅

Ada & luna New Year’s Eve 2025 The year we learned to navigate the stars 🪐✨φ